Check your MCP setup
Paste your MCP client config: claude_desktop_config.json (Claude Desktop), mcp.json (Cursor, VS Code), .mcp.json (Claude Code) or similar. You get a grade and a list of fixes.
Nothing is uploaded. The check runs entirely in your browser. Your config is not sent, stored or logged anywhere; this page makes no network requests with it. Secrets found in it are shown masked.
This check needs JavaScript. It runs only on your device.
Paste your config
What it looks for
- High: a filesystem server rooted at
/,~, your home folder or a whole drive. - High: a docker server with
--privileged, powerful capabilities, the Docker socket or a very broad host mount. - High: a remote server over plain
http://(not localhost). - Medium: an
npx,bunx,uvx,pnpm dlxpackage with no exact version, or a moving tag like@latest/:latest(docker images too). - Low: a secret pasted in plain text in
env, headers or arguments; a writable host folder mounted into docker. - Servers found on the board show their grade and how many risky tools have no safety label.
How your grade is worked out
- Start at 100.
- −15 for each high finding.
- −10 for each medium finding.
- −5 for each low finding (one per server for pasted secrets).
- −1 for each risky tool without a safety label in servers found on the board, at most −20.
- Never below 0. A ≥ 90 · B 80–89 · C 70–79 · D 60–69 · F < 60, the same bands as the board.
Board matches use the npm or PyPI package name published in each graded repo at the commit checked, ghcr.io image names and GitHub links. Local scripts can't be matched.
Want a written review of your agent's tools and permissions?
48-hour written review. Which tools your agent can call, what each one can change, which need a confirmation step, and what to lock down first. Send a message to start.